A Multi-Fidelity Approach to Testing and Evaluation of AI-Enabled Systems

Abstract As Artificial Intelligence (AI) and Machine Learning (ML) technologies advance, the need for methods to test them also increases. Several methods for testing ML and AI systems have recently been developed by the ML/AI and software development communities. These methods assume that the team that is developing the system is also responsible for testing the system, therefore, they have access to the datasets on which the ML models were trained, and the knowledge of the environment in which the system is expected to operate. However, this is not true in certain situations such as in the Department of Defense (DoD) acquisition where the systems are developed by other organizations, due to which, existing methods for Test and Evaluation (T&E) of AI-enabled systems are not adequate for DoD acquisition. To address this gap we propose a multi-fidelity approach for testing and evaluation that consists of (i) a representation of the model space with dimensions along which different fidelities of models can be developed, and (ii) a method to integrate multiple fidelities for continuous T&E of AI-enabled systems. The approach is illustrated using an example of a visual perception system in an autonomous vehicle (AV) use case, where a simulation space across different fidelities is constructed to test how well the system meets the listed requirements. A model space is first identified, in which models are characterized for their cost and performance. A method to generate test plans is then devised to maximize the utility across the span of given system requirements. We illustrate how the proposed approach can be used to develop test combinations that minimize the cost and maximize the utility under a set of system requirements to be tested.